Slop Detector

Slop Detector

Scans a draft for ten fixed signatures of AI-generated explainer slop: meta-framing openers ("In this post"), list-carried argument, nominalization clusters, generic examples with no first-person texture, prompt-residue phrases ("Let's break this down"), buzzword stuffing, outline-shaped paragraphs, hedge clusters, and flattened uncertainty. Each signature has a concrete detection rule and is reported as clean or flagged with the offending span quoted. It does not measure sentence-level readability; that is readability-check's job.

When to Use

• A draft "feels generic" even after a voice pass has been applied.
• Reviewing any long-form output (report, essay, memo, forecast write-up) before it is delivered.
• The user mentions slop, AI-written, generic, template, meta-framing, zombie nouns, prompt residue, or outline-shaped prose.
• A document is heavy on "First / Second / Third" lists, "To summarize" closers, or "a company might" examples.

Related skills: run this before readability-check; slop is a content problem, readability a structural one, and fixing slop often changes the sentences the readability pass would have scored.

The 10 signatures

Fixed list. Each is either clean or flagged with the offending span.

Procedure

Slop scan draft D:
- [ ] Step 1: For each signature, run the detection rule
- [ ] Step 2: Mark each signature clean | flagged (with quote)
- [ ] Step 3: Tier-1 signatures: S1, S2, S6 (generic framing + prompt residue)
- [ ] Step 4: Tier-2 signatures: S3, S4, S5, S7, S8, S9
- [ ] Step 5: Emit the slop-signatures subsection with each labelled clean/flagged

S3 nominalization scoring

Count suffix hits (-ation, -ity, -ment, -ence, -ness, -ance) on abstract nouns per 100 words. >3 = flag. Example: "provides analysis of" is nominalized; "analyzes" is active.

S4 generic-example rule

Flag an example if it uses only generic pronouns or nouns without a specific anchor:

• "A company might use this": flag.
• "At Google in 2024, Chen et al. used this": clean.

S7 outline-shape rule

Parse paragraphs; count those with the shape:

• Sentence 1: topic statement
• Sentences 2-4: three supporting sentences
• Last sentence: transition

60% of paragraphs following this shape means the draft reads like an outline expanded by a model.

S8 hedge-cluster rule

Count epistemic-weakness hedges (phrases that weaken a claim without adding information). Two or more within 50 words is a cluster. A single precise hedge ("the sample is 12 firms, so treat the rate as indicative") is not a hedge for this purpose; it is a caveat and belongs in the text.

Worked example

Draft fragment:

In this post, we'll explore why RAG beats fine-tuning.

First, let's define RAG. It's a technique where models retrieve documents before generating. A company might use RAG for their customer service chatbot.

Second, fine-tuning involves training. A team might fine-tune to adapt style.

Third, RAG has benefits. Fine-tuning has drawbacks. It could be argued that hybrid works.

To summarize, both approaches have merit.

Detections:

• S1: flagged ("In this post, we'll explore").
• S2: flagged (argument carried by the "First / Second / Third" list-in-prose).
• S3: zombie-noun check: "technique", "documents", "benefits", "drawbacks"; borderline, not flagged yet.
• S4: flagged ("A company might use RAG", "a team might fine-tune"); no specifics.
• S5: clean (has "we").
• S6: flagged ("To summarize").
• S7: flagged (each paragraph: topic, supporting, transition).
• S8: weakness hedges: "It could be argued"; just one, not a cluster (yet).
• S9: buzzwords: "explore" is close; not flagged (one term).
• S10: skipped (no notes).

Output: 5 signatures flagged (S1, S2, S4, S6, S7). Tier-1: S1, S2, S6, so 3 tier-1 slop violations.

Output format

Slop signatures:
- S1 meta-framing opener: flagged - "In this post, we'll explore"
- S2 list-carrying-argument: flagged - "First ... Second ... Third"
- S3 zombie nouns: clean (2.1 per 100 words)
- ...
Tier-1 violations: 3 (S1, S2, S6). Tier-2 violations: 2 (S4, S7).

Guardrails

1. Each signature has a concrete detection rule. No "feels like slop."
2. Quote the offending span. Do not just say "S1 triggered"; quote the opener.
3. Signatures are additive, not exclusive. A draft can trip 8 signatures and still be revisable; the calling workflow decides the no-go threshold (a common rule: any tier-1 hit blocks delivery).
4. S10 requires the author's working notes; skip quietly if absent.
5. If a separate hedge-detection pass has already run, take its cluster count as the S8 input rather than scanning twice.
6. S5 (no first person) applies to reflective essays only. How-to and methodology documents may legitimately lack "I".

Quick Reference

• 10 fixed signatures, deterministic detection.
• Tier-1: S1, S2, S6. Tier-2: the rest.
• Output is a per-signature clean/flagged list with quotes, then the tier counts.

Verification

Run the scan on the worked example above: it must reproduce exactly five flags (S1, S2, S4, S6, S7) with the quoted spans. A scan that flags fewer is skipping rules; one that flags S3 or S8 on this fragment is not applying the thresholds.